Alibaba’s Qwen team published open weights for two Qwen3.8 models. Only the smaller 27-billion-parameter model carries an Apache 2.0 license; the 2.4-trillion-parameter flagship ships under a custom license with commercial revenue conditions.
Key Takeaways
- Qwen3.8-27B ships under Apache 2.0 with 27 billion parameters, native image and video input, and a 262,144-token context window.
- Qwen3.8-2.4T-A95B uses a separate Qwen3.8-Max License rather than Apache 2.0, despite arriving in the same open weights announcement.
- The Qwen3.8-Max License makes firms running an AI service business above $50 million in yearly revenue obtain a separate license from Qwen.
- Qwen reports 61.7% on SWE-bench Pro for the 27B model, a self-published figure that outside labs have not replicated.
- Both models are live on Hugging Face and ModelScope, with a hosted million-token version planned for Alibaba’s Qwen Cloud.
What We Know?
Alibaba’s Qwen team posted weights for two Qwen3.8 models to Hugging Face and ModelScope, closing a commitment the team made when it introduced Qwen3.8-Max. One release is a dense multimodal model sized for local hardware. The other is a Mixture of Experts flagship that needs a data center.
Qwen3.8-27B processes text, images and video in a single model, runs a 262,144-token context window natively, and stretches toward 1 million tokens through YaRN, a context extension method. Its thinking mode is on by default and can be switched off per request, with reasoning depth set to xhigh, medium or low. The model card credits the release with “substantial gains across coding, professional work, research, and long-horizon agentic tasks.“
The Qwen3.8-2.4T-A95B model carries 2.4 trillion total parameters and activates 95 billion per request across 512 experts. Its card calls it “the first Qwen-Max-class model to open release.”
We promised open weights for Qwen3.8. Now, time to meet them! 🎉
— Qwen (@Alibaba_Qwen) August 14, 2026
⚡ Qwen3.8-27B:
– A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows.
– 262K native context, easily extendable to 1M… pic.twitter.com/QuN8oWkG4C
Two models, two very different licenses
The licensing splits at the top of the lineup, and the split is easy to miss in the announcement. Qwen3.8-27B is Apache 2.0, the permissive standard developers expect from an open weights drop. Qwen3.8-2.4T-A95B is not.
The flagship ships under a document titled the Qwen3.8-Max License. It grants rights to “use, copy, modify, merge, publish, distribute, sublicense, sell, deploy, host, fine-tune, and create derivative works,” then attaches two conditions Apache 2.0 does not carry.
Products serving more than 100 million monthly active users or $20 million in monthly revenue must display the model name prominently in the user interface. Any company operating a Model as a Service or AI Work Assistant business whose aggregate revenue tops $50 million across any consecutive 12 months “shall obtain a separate license from Qwen before Using the Software.“
That second clause lands squarely on inference providers and agent startups, the exact group most likely to want a Max class model. The license draws its own boundary around internal deployment, defining Model as a Service by third-party access and excluding internal use that never exposes the model outward.
What the benchmarks claim?
Qwen published its own scores for both models. No independent lab has reproduced them.
- Qwen3.8-27B: 61.7% on SWE-bench Pro, 84.3% on OSWorld-Verified, 64.8% on WebArena-Verified.
- Qwen3.8-27B: 94.6% on MathVision with chain-of-thought prompting.
- Qwen3.8-2.4T-A95B: 67.7 on SWE-bench Pro, 93.0 on PaperBench, 92.6 on GPQA Diamond.
The 27B model is the one most developers will actually run. A dense model at that size fits on a single high-memory accelerator, putting it in the same practical bracket as Zhipu AI’s GLM-4.5 open release and far below what the 2.4 trillion-parameter flagship demands in AI accelerator hardware. On multimodality, it puts Alibaba alongside SenseTime’s SenseNova U1 and ahead of text-only rivals including DeepSeek’s V4 line.
What We Don’t Know?
A claim circulated after the release that the flagship license barred downloads from the United States, European Union, United Kingdom and South Korea. The published license text on Hugging Face contains no territory or jurisdiction clause. The document supports the revenue and attribution conditions above; it does not support a geographic ban.
Open questions Alibaba has not addressed:
- Why the two models carry different licenses when both were announced as open weights?
- Whether the Qwen3.8-Max License terms will apply to future Max-class releases?
- What the review process looks like for a company that must request the separate commercial license?
- When the hosted million-token version reaches Qwen Cloud?
Teams that already pulled the 2.4T weights should check today whether their organization crosses either revenue threshold before shipping anything built on the model. The 27B release under Apache 2.0 carries no equivalent condition, which makes it the safer starting point for a commercial product while legal review runs.
SQ Magazine’s Takeaway
The release reads as two decisions wearing one announcement. Apache 2.0 on the 27B model buys developer goodwill and local adoption at low strategic cost, since a 27 billion-parameter model competes with other open releases rather than with Alibaba’s paid tier. The custom license on the 2.4T flagship protects the commercial layer, keeping large inference resellers inside a negotiation instead of a download. Calling both releases open weights is accurate. Calling both of them Apache 2.0, as several early writeups did, is not.
What’s next depends on how quickly independent evaluations land. Outside benchmark runs from groups such as LMArena or academic labs would test whether the 61.7% SWE-bench Pro figure holds under conditions Qwen did not control. Legal teams at inference providers will want a reading on the $50 million threshold before production deployment, particularly on whether the revenue test counts total company revenue or the AI service line alone. The license text does not say.
